Scheduler Optimization: Pre-allocate vectors during buffering checks#1725
Scheduler Optimization: Pre-allocate vectors during buffering checks#1725apfitzge merged 3 commits intoanza-xyz:masterfrom
Conversation
| @@ -531,9 +531,9 @@ impl SchedulerController { | |||
| let mut num_dropped_on_capacity: usize = 0; | |||
| let mut num_buffered: usize = 0; | |||
| for (((packet, transaction), fee_budget_limits), _) in arc_packets | |||
| let mut arc_packets = Vec::with_capacity(CHUNK_SIZE); | ||
| let mut transactions = Vec::with_capacity(CHUNK_SIZE); | ||
| let mut fee_budget_limits_vec = Vec::with_capacity(CHUNK_SIZE); |
There was a problem hiding this comment.
nit: using smallvec further reduces allocation entirely per each buffer_packets() invocation.
There was a problem hiding this comment.
I opted for arrayvec, since it fully guarantees no heap-allocation. I think it's also slightly faster because there's no need to check for overflow onto the heap when we drain them.
There was a problem hiding this comment.
yeah, by principle, it'll be slightly faster.
unlike smallvec, arrayvec will panic!() if we're about to .push() more than CHUNK_SIZE. However, that seems to be guaranteed not to happen.
There was a problem hiding this comment.
Also, the trio of these arrayvec will consume ~36 kb of stack in total, according to rust-analyzer:
// size = 1032 (0x408), align = 0x8
let mut arc_packets: ArrayVec<Arc<ImmutableDeserializedPacket>, 128>
// size = 31752 (0x7C08), align = 0x8
let mut transactions: ArrayVec<SanitizedTransaction, 128>
// size = 4104 (0x1008), align = 0x8
let mut fee_budget_limits_vec: ArrayVec<FeeBudgetLimits, 128>That again will be okay, considering linux's default thread stack size. (8 MiB)
| @@ -479,14 +480,14 @@ impl SchedulerController { | |||
|
|
|||
| const CHUNK_SIZE: usize = 128; | |||
| let lock_results: [_; CHUNK_SIZE] = core::array::from_fn(|_| Ok(())); | |||
There was a problem hiding this comment.
(off topic) this can be rewritten with inline const once we update to rust 1.79.
3 participants
Problem
Summary of Changes
drainto avoid lifetime issues with the vecs continuing to live onPulled out of #1663 (thanks @t-nelson)
Fixes #